This commit is a follow-up to https://github.com/facebook/react/pull/15604, which explains more of the rationale behind moving React Native to path-based imports and the work needed in the React repository. In that linked PR, the generated renderers were updated but not the shims; this commit updates the shims.
The problem is that FB needs a different copy of the built renderers than the OSS versions so we need a way for FB code to import different modules than in OSS. This was previously done with Haste, but with the removal of Haste from RN, we need another mechanism. Talking with cpojer, we are using a `.fb.js` extension that Metro can be configured to prioritize over `.js`.
This commit generates FB's renderers with the `.fb.js` extension and OSS renderers with just `.js`. This way, FB can internally configure Metro to use the `.fb.js` implementations and OSS will use the `.js` ones, letting us swap out which implementation gets bundled.
Test Plan: Generated the renderers and shims with `yarn build` and then verified that the generated shims don't contain any Haste-style imports. Copied the renderers and shims into RN manually and launched the RNTester app to verify it loads end-to-end. Added `.fb.js` to the extensions in `metro.config.js` and verified that the FB-specific bundles loaded.
The previous naming scheme used the name of the resulting bundle file.
However, there are cases where multiple bundles have the same filename.
This meant whichever bundle finishes last overwrites the previous ones
with the same name.
The updated naming scheme is `bundle-sizes-<CI_NODE_INDEX>.json`.
Instead of generating a separate info file per bundle, it now creates
one per process.
* Write size info to separate file per bundle
`bundle-sizes.json` contains the combined size information for every
build. This makes it easier to store and process, but it prevents us
from parallelizing the build script, because each process would need to
write to the same file.
So I've updated the Rollup script to output individual files per build.
A downstream CI job consolidates them into a single file.
I have not parallelized the Rollup script yet. I'll do that next.
* Parallelize the build script
Uses CircleCI's `parallelism` config option to spin up multiple build
processes.
Removes `--extract-errors` argument from CI build script command.
Instead, the author is expected to run `yarn extract-errors` locally
or manually edit the error code map.
The lint rule should be sufficient to catch unminified errors, but
as an extra precaution, I added a post-build step that greps the
production bundles. The post-build step works even if someone disables
the lint rule for a specific line or file.
* Lint rule for unminified errors
Add a lint rule that fails if an invariant message is not part of the
error code map.
The goal is to be more disciplined about adding and modifiying
production error codes. Error codes should be consistent across releases
even if their wording changes, for continuity in logs.
Currently, error codes are added to the error code map via an automated
script that runs right before release. The problem with this approach is
that if someone modifies an error message in the source, but neglects to
modify the corresponding message in the error code map, then the message
will be assigned a new error code, instead of reusing the existing one.
Because the error extraction script only runs before a release, people
rarely modify the error code map in practice. By moving the extraction
step to the PR stage, it forces the author to consider whether the
message should be assigned a new error code. It also allows the reviewer
to review the changes.
The trade off is that it requires more effort and context to land new
error messages, or to modify existing ones, particular for new
contributors who are not familiar with our processes.
Since we already expect users to lint their code, I would argue the
additional burden is marginal. Even if they forget to run the lint
command locally, they will get quick feedback from the CI lint job,
which typically finishes within 2-3 minutes.
* Add unreleased error messages to map
* [react-native] Use path-based imports instead of Haste for the RN renderer
To move React Native to standard path-based imports instead of Haste, the RN renderer that is generated from the code in this repo needs to use path-based imports as well since the generated code is vendored by RN. This commit makes it so the interface between the generated renderers and RN does not rely on Haste and instead uses a private interface explicitly defined by RN. This inverts control of the abstraction so that RN decides the internals to export rather than React deciding what to import.
On RN's side, a new module named `react-native/Libraries/ReactPrivate/ReactNativePrivateInterface` explicitly exports the modules used by the renderers in this repo. (There is also a private module for InitializeCore so that we can import it just for the side effects.) On React's side, the various renderer modules access RN internals through the explicit private interface.
The Rollup configuration becomes slimmer since the only external package is now `react-native`, and the individual modules are instead listed out in `ReactNativePrivateInterface`.
Task description: https://github.com/facebook/react-native/issues/24770
Sister RN PR (needs to land before this one): https://github.com/facebook/react-native/pull/24782
Test Plan: Ran unit tests and Flow in this repo. Generated the renderers and manually copied them over to the RN repo. Ran the RN tests and launched the RNTester app.
* Access natively defined "nativeFabricUIManager" instead of importing it
Some places in the Fabric renderers access `nativeFabricUIManager` (a natively defined global) instead of importing UIManager. While this is coupling across repos that depends on the timing of events, it is necessary until we have a way to defer top-level imports to run after `nativeFabricUIManager` is defined. So for consistency we use `nativeFabricUIManager` everywhere (see the comment in https://github.com/facebook/react/pull/15604#pullrequestreview-236842223 for more context).
Updates the CircleCI config to use the workflows features to run jobs in
parallel, instead of the `parallelism` option. This change alone doesn't
improve the overall build time much, since almost all of the total time
is spent running the Rollup script, which runs entirely sequentially.
But it does improve reporting, and should make it easier to add
additional parallel jobs in the future.
* Add a stub for React Fresh Babel plugin package
* Move ReactFresh-test into ReactFresh top level directory
* Add a stub for React Fresh Runtime entry point
* Extract Fresh runtime from tests into its entry point
* s/flushPassiveEffects/unstable_flushWithoutYielding
a first crack at flushing the scheduler manually from inside act(). uses unstable_flushWithoutYielding(). The tests that changed, mostly replaced toFlushAndYield(...) with toHaveYielded(). For some tests that tested the state of the tree before flushing effects (but still after updates), I replaced act() with bacthedUpdates().
* ugh lint
* pass build, flushPassiveEffects returns nothing now
* pass test-fire
* flush all work (not just effects), add a compatibility mode
of note, unstable_flushWithoutYielding now returns a boolean much like flushPassiveEffects
* umd build for scheduler/unstable_mock, pass the fixture with it
* add a comment to Shcduler.umd.js for why we're exporting unstable_flushWithoutYielding
* run testsutilsact tests in both sync/concurrent modes
* augh lint
* use a feature flag for the missing mock scheduler warning
I also tried writing a test for it, but couldn't get the scheduler to unmock. included the failing test.
* Update ReactTestUtilsAct-test.js
- pass the mock scheduler warning test,
- rewrite some tests to use Scheduler.yieldValue
- structure concurrent/legacy suites neatly
* pass failing tests in batchedmode-test
* fix pretty/lint/import errors
* pass test-build
* nit: pull .create(null) out of the act() call
If React finishes rendering a tree, delays committing it (e.g.
Suspense), then subsequently starts over or renders a new tree, the
pending tree is no longer valid. That's because rendering a new work-in
progress mutates the old one in place.
The current structure of the work loop makes this hard to reason about
because, although `renderRoot` and `commitRoot` are separate functions,
they can't be interleaved. If they are interleaved by accident, it
either results in inconsistent render output or invariant violations
that are hard to debug.
This commit adds an invariant that throws if the new tree is the same as
the old one. This won't prevent all bugs of this class, but it should
catch the most common kind.
To implement the invariant, I store the finished tree on a field on the
root. We already had a field for this, but it was only being used for
the unstable `createBatch` feature.
A more rigorous way to address this type of problem could be to unify
`renderRoot` and `commitRoot` into a single function, so that it's
harder to accidentally interleave the two phases. I plan to do something
like this in a follow-up.
* [React Native] Inline calls to FabricUIManager in shared code
* Call global.nativeFabricUIManager directly as short term fix
* Add flow types
* Add nativeFabricUIManager global to eslint config
* Adding eslint global to bundle validation script
The React Native build does not minify error messages in production,
but it still needs to run the error messages transform to compile
`invariant` calls to `ReactError`. To do this, I added a `noMinify`
option to the Babel plugin. I also renamed it from
`minify-error-messages` to the more generic `transform-error-messages`.
* [React Native] Add tests to paper renderer for measure, measureLayout
* [React Native] measure calls will now call FabricUIManager
The Fabric renderer was previously calling the paper UIManager's measure calls and passing the react tag. This PR changes the renderer to now call FabricUIManager passing the node instead.
One of the parts of this that feels more controversial is making NativeMethodsMixin and ReactNative.NativeComponent warn when calling measureLayout in Fabric. As Seb and I decided in https://github.com/facebook/react/pull/15126, it doesn't make sense for a component created with one of these methods to require a native ref but not work the other way around. For example: a.measureLayout(b) might work but b.measureLayout(a) wouldn't. We figure we should keep these consistent and continue migrating things off of NativeMethodsMixin and NativeComponent.
If this becomes problematic for the Fabric rollout then we should revisit this.
* Fixing Flow
* Add FabricUIManager to externals for paper renderer
* import * as FabricUIManager from 'FabricUIManager';
* Update tests
* Shouldn't have removed UIManager import
* Update with the new tests
* Store FB bundles as CI artifacts
Updates the Circle CI config to store Facebook bundles as build
artifacts. We already do this for our npm packages.
* Might as well store everything in build/
* Store build directory as a tarball
So it's easy to download
* Rewrite ReactFiberScheduler
Adds a new implementation of ReactFiberScheduler behind a feature flag.
We will maintain both implementations in parallel until the new one
is proven stable enough to replace the old one.
The main difference between the implementations is that the new one is
integrated with the Scheduler package's priority levels.
* Conditionally add fields to FiberRoot
Some fields only used by the old scheduler, and some by the new.
* Add separate build that enables new scheduler
* Re-enable skipped test
If synchronous updates are scheduled by a passive effect, that work
should be flushed synchronously, even if flushPassiveEffects is
called inside batchedUpdates.
* Passive effects have same priority as render
* Revert ability to cancel the current callback
React doesn't need this anyway because it never schedules callbacks if
it's already rendering.
* Revert change to FiberDebugPerf
Turns out this isn't neccessary.
* Fix ReactFiberScheduler dead code elimination
Should initialize to nothing, then assign the exports conditionally,
instead of initializing to the old exports and then reassigning to the
new ones.
* Don't yield before commit during sync error retry
* Call Scheduler.flushAll unconditionally in tests
Instead of wrapping in enableNewScheduler flag.
This took a while, but I'm happy I went through it. Some key moments - recursively flushing effects, flushing microtasks on each async turn, and my team's uncompromising philosophy on code reuse. Really happy with this. I still want to expand test coverage, and I have some more small related todos, but this is good to land. On to the next one.
Soundtrack to landing this - https://open.spotify.com/track/0MF8I8OUo8kytiOo8aSHYq?si=gSWqUheKQbiQDXzptCXHTg
* hacked up act(async () => {...})
* move stuff around
* merge changes
* abstract .act warnings and stuff. all renderers. pass all tests.
* move testutils.act back into testutils
* move into scheduler, rename some bits
* smaller bundle
* a comment for why we don't do typeof === 'function'
* fix test
* pass tests - fire, prod
* lose actContainerElement
* tighter
* write a test for TestRenderer
it's an odd one, because not only does sync act not flush effects correctly, but the async one does (wut). verified it's fine with the dom version.
* lint
* rewrote to move flushing logic closer to the renderer
the scheduler's `flushPassiveEffects` didn't work as expected for the test renderer, so I decided to go back to the hack (rendering a dumb container) This also makes reactdom not as heavy (by a few bytes, but still).
* move it around so the delta isn't too bad
* cleanups
fix promise chaining
propagate errors correctly
test for thenable the 'right' way
more tests!
tidier!
ponies!
* Stray comment
* recursively flush effects
* fixed tests
* lint, move noop.act into react-reconciler
* microtasks when checking if called, s/called/calledLog, cleanup
* pass fb lint
we could have globally changed our eslint config to assume Promise is available, but that means we expect a promise polyfill on the page, and we don't yet. this code is triggered only in jest anyway, and we're fairly certain Promise will be available there. hence, the once-off disable for the check
* shorter timers, fix a test, test for Promise
* use global.Promise for existence check
* flush microtasks
* a version that works in browsers (that support postMessage)
I also added a sanity fixture inside fixtures/dom/ mostly for me.
* hoist flushEffectsAndMicroTasks
* pull out tick logic from ReactFiberScheduler
* fix await act (...sync) hanging
- fix a hang when awaiting sync logic
- a better async/await test for test renderer
* feedback changes
- use node's setImmediate if available
- a warning if MessageChannel isn't available
- rename some functions
* pass lint/flow checks (without requiring a Promise polyfill/exclusion)
* prettier
the prettiest, even.
* use globalPromise for the missed await warning
* __DEV__ check for didWarnAboutMessageChannel
* thenables and callbacks instead of promises, pass flow/lint
* tinier. better.
- pulled most bits out of FiberScheduler
- actedUpdates uses callbacks now
* pass build validation
* augh prettier
* golfing 7 more chars
* Test that effects are not flushed without also flushing microtasks
* export doesHavePendingPassiveEffects, nits
* createAct()
* dead code
* missed in merge?
* lose the preflushing bits
* ugh prettier
* removed `actedUpdates()`, created shared/actingUpdatesScopeDepth
* rearrange imports so builds work, remove the hack versions of flushPassiveEffects
* represent actingUpdatesScopeDepth as a tuple [number]
* use a shared flag on React.__SECRET...
* remove createAct, setup act for all relevant renderers
* review feedback
shared/enqueueTask
import ReactSharedInternals from 'shared/ReactSharedInternals';
simpler act() internals
ReactSharedInternals.ReactShouldWarnActingUpdates
* move act() implementation into createReactNoop
* warnIfNotCurrentlyActingUpdatesInDev condition check order
Adds a feature flag `enableNewScheduler` that toggles between two
implementations of ReactFiberScheduler. This will let us land changes in
master while preserving the ability to quickly rollback.
Ideally this will be a short-lived fork. Once we've tested the new
scheduler for a week or so without issues, we will get rid of it. Until
then, we'll need to maintain two parallel implementations and run tests
against both of them. We rarely land changes to ReactFiberScheduler, so
I don't expect this will be a huge burden.
This commit does not implement anything new. The flag is still off and
tests run against the existing implementation.
Use `yarn test-new-scheduler` to run tests against the new one.
* Transform invariant to custom error type
This transforms calls to the invariant module:
```js
invariant(condition, 'A %s message that contains %s', adj, noun);
```
Into throw statements:
```js
if (!condition) {
if (__DEV__) {
throw ReactError(`A ${adj} message that contains ${noun}`);
} else {
throw ReactErrorProd(ERR_CODE, adj, noun);
}
}
```
The only thing ReactError does is return an error whose name is set
to "Invariant Violation" to match the existing behavior.
ReactErrorProd is a special version used in production that throws
a minified error code, with a link to see to expanded form. This
replaces the reactProdInvariant module.
As a next step, I would like to replace our use of the invariant module
for user facing errors by transforming normal Error constructors to
ReactError and ReactErrorProd. (We can continue using invariant for
internal React errors that are meant to be unreachable, which was the
original purpose of invariant.)
* Use numbers instead of strings for error codes
* Use arguments instead of an array
I wasn't sure about this part so I asked Sebastian, and his rationale
was that using arguments will make ReactErrorProd slightly slower, but
using an array will likely make all the functions that throw slightly
slower to compile, so it's hard to say which way is better. But since
ReactErrorProd is in an error path, and fewer bytes is generally better,
no array is good.
* Casing nit
* Add command to run tests in persistent mode
* Convert Suspense fuzz tester to use noop renderer
So we can run it in persistent mode, too.
* Don't mutate stateNode in appendAllChildren
We can't mutate the stateNode in appendAllChildren because the children
could be current.
This is a bit weird because now the child that we append is different
from the one on the fiber stateNode. I think this makes conceptual
sense, but I suspect this likely breaks an assumption in Fabric.
With this approach, we no longer need to clone to unhide the children,
so I removed those host config methods.
Fixes bug surfaced by fuzz tester. (The test case that failed was the
one that's already hard coded.)
* In persistent mode, disable test that reads a ref
Refs behave differently in persistent mode. I added a TODO to write
a persistent mode version of this test.
* Run persistent mode tests in CI
* test-persistent should skip files without noop
If a file doesn't reference react-noop-renderer, we shouldn't bother
running it in persistent mode, since the results will be identical to
the normal test run.
* Remove module constructor from placeholder tests
We don't need this now that we have the ability to run any test file in
either mutation or persistent mode.
* Revert "test-persistent should skip files without noop"
Seb objected to adding shelljs as a dep and I'm too lazy to worry about
Windows support so whatever I'll just revert this.
* Delete duplicate file
* Swap expect(ReactNoop) for expect(Scheduler)
In the previous commits, I upgraded our custom Jest matchers for the
noop and test renderers to use Scheduler under the hood.
Now that all these matchers are using Scheduler, we can drop
support for passing ReactNoop and test roots and always pass
Scheduler directly.
* Externalize Scheduler in noop and test bundles
I also noticed we don't need to regenerator runtime in noop anymore.
* Replace test renderer's fake Scheduler implementation with mock build
The test renderer has its own mock implementation of the Scheduler
interface, with the ability to partially render work in tests. Now that
this functionality has been lifted into a proper mock Scheduler build,
we can use that instead.
* Fix Profiler tests in prod
* Replace noop's fake Scheduler implementation with mock Scheduler build
The noop renderer has its own mock implementation of the Scheduler
interface, with the ability to partially render work in tests. Now that
this functionality has been lifted into a proper mock Scheduler build,
we can use that instead.
Most of the existing noop tests were unaffected, but I did have to make
some changes. The biggest one involved passive effects: previously, they
were scheduled on a separate queue from the queue that handles
rendering. After this change, both rendering and effects are scheduled
in the Scheduler queue. I think this is a better approach because tests
no longer have to worry about the difference; if you call `flushAll`,
all the work is flushed, both rendering and effects. But for those few
tests that do care to flush the rendering without the effects, that's
still possible using the `yieldValue` API.
Follow-up: Do the same for test renderer.
* Fix import to scheduler/unstable_mock
* Add new mock build of Scheduler with flush, yield API
Test environments need a way to take control of the Scheduler queue and
incrementally flush work. Our current tests accomplish this either using
dynamic injection, or by using Jest's fake timers feature. Both of these
options are fragile and rely too much on implementation details.
In this new approach, we have a separate build of Scheduler that is
specifically designed for test environments. We mock the default
implementation like we would any other module; in our case, via Jest.
This special build has methods like `flushAll` and `yieldValue` that
control when work is flushed. These methods are based on equivalent
methods we've been using to write incremental React tests. Eventually
we may want to migrate the React tests to interact with the mock
Scheduler directly, instead of going through the host config like we
currently do.
For now, I'm using our custom static injection infrastructure to create
the two builds of Scheduler — a default build for DOM (which falls back
to a naive timer based implementation), and the new mock build. I did it
this way because it allows me to share most of the implementation, which
isn't specific to a host environment — e.g. everything related to the
priority queue. It may be better to duplicate the shared code instead,
especially considering that future environments (like React Native) may
have entirely forked implementations. I'd prefer to wait until the
implementation stabilizes before worrying about that, but I'm open to
changing this now if we decide it's important enough.
* Mock Scheduler in bundle tests, too
* Remove special case by making regex more restrictive
* Throw in tests if work is done before emptying log
Test renderer already does this. Makes it harder to miss unexpected
behavior by forcing you to assert on every logged value.
* Convert ReactNoop tests to use jest matchers
The matchers warn if work is flushed while the log is empty. This is
the pattern we already follow for test renderer. I've used the same APIs
as test renderer, so it should be easy to switch between the two.
* Improve release script process documentation
* Improved pre-publish instructions/message based on feedback
* Added reminder to attach build artifacts to GitHub release
